Ground object identification method and device for disaster response and electronic equipment

Abstract

The invention discloses a surface feature recognition method and device for disaster response and electronic equipment, and relates to the technical field of artificial intelligence, and the surface feature recognition method comprises the steps: obtaining a remote sensing image of a disaster area, carrying out the preprocessing of the remote sensing image, obtaining a target remote sensing image, carrying out the marking of the target remote sensing image, and generating an initial tag map, and iterating a preset network structure based on the initial label map until the number of iterations reaches a preset number of iterations to obtain a ground feature recognition model, and recognizing the target remote sensing image by using the ground feature recognition model to obtain a ground feature classification map of the disaster region. According to the invention, the technical problem of poor disaster response timeliness caused by the fact that ground object identification cannot be rapidly carried out in the prior art is solved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically, to a method, apparatus, and electronic device for identifying ground features in disaster response. Background Technology

[0002] In disaster emergency response, rapid disaster assessment using remote sensing imagery has become a primary technical means. With the rapid development of deep learning technology, especially convolutional neural networks, the use of convolutional neural networks for ground feature identification in remote sensing imagery tasks can improve accuracy and efficiency.

[0003] In related technologies, fully supervised semantic segmentation networks are generally used to perform pixel-level classification of remote sensing images before and after disasters, thereby extracting information on changes in ground features. However, the above-mentioned fully supervised deep learning methods are highly dependent on large-scale, high-quality, and finely labeled datasets. In real, sudden disaster response scenarios, they face the following challenges: (1) Time urgency: Time is precious after a disaster, and large-scale manual labeling requires huge human and material resources; (2) Data scarcity: The characteristics (lighting, image quality, and ground feature morphology) of each disaster are different, and there is a lack of readily available, fully matching labeled datasets.

[0004] Therefore, although the aforementioned fully supervised deep learning methods have high accuracy, when applied to rapid disaster response, they rely heavily on large-scale finely labeled samples, resulting in long model training and preparation cycles and slow response speeds, which cannot meet the timeliness requirements of emergency scenarios.

[0005] There is currently no effective solution to the above problems. Summary of the Invention

[0006] This invention provides a method, apparatus, and electronic device for identifying ground features in disaster response, to at least solve the technical problem in related technologies where ground feature identification is not possible quickly, resulting in poor timeliness of disaster response.

[0007] According to one aspect of the present invention, a method for land cover identification in disaster response is provided, comprising: acquiring remote sensing images of a disaster area and preprocessing the remote sensing images to obtain target remote sensing images; labeling the target remote sensing images to generate an initial label map, wherein the initial label map is used to characterize the initial label of each pixel on the target remote sensing image, and the initial label is used to characterize the land cover category of the pixel; iterating a preset network structure based on the initial label map until the number of iterations reaches a preset number of iterations to obtain a land cover identification model, wherein the iteration of the preset network structure is achieved by continuously iterating the initial label map; and using the land cover identification model to identify the target remote sensing images to obtain a land cover classification map of the disaster area, wherein the land cover classification map is used to display the categories and location ranges of land cover in the disaster area.

[0008] Further, the step of preprocessing the remote sensing image to obtain the target remote sensing image includes: performing radiometric calibration on the remote sensing image to obtain a first remote sensing image; performing atmospheric correction on the first remote sensing image to obtain a second remote sensing image; and performing orthorectification on the second remote sensing image to obtain the target remote sensing image.

[0009] Further, the step of annotating the target remote sensing image and generating an initial label map includes: segmenting the target remote sensing image based on different scale parameters to obtain an image object set corresponding to each scale, wherein the image objects in the image object sets corresponding to different scales are of different sizes; for each image object set, selecting a preset number of image objects from the image object set for annotation to obtain the annotation category of the image object; based on all image objects annotated with annotation categories, using a nearest neighbor classification strategy to classify the remaining image objects that are not annotated with annotation categories to generate an initial label map, wherein the initial label map is used to represent the initial category of each image object.

[0010] Further, based on all image objects labeled with labeled categories, the step of classifying the remaining image objects without labeled categories using a nearest neighbor classification strategy to generate an initial label map includes: extracting the feature vector of each image object, wherein the feature vector includes at least one of the following: spectral feature vector, texture feature vector, and shape feature vector; for each remaining image object, calculating a preset distance between the feature vector of the remaining image object and the feature vector of each preset image object, wherein the preset image object refers to the image object labeled with labeled categories; assigning the labeled category of the preset image object indicated by the minimum preset distance to the remaining image object to obtain the initial category of the remaining image object; and generating an initial label map based on the labeled categories of all preset image objects and the initial categories of all remaining image objects.

[0011] Furthermore, before iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the land cover recognition model, the process includes: constructing a backbone network, wherein the backbone network includes: multiple convolutional structures, each of which corresponds to a deformable convolutional structure; the backbone network is used to extract feature maps from the input remote sensing image; constructing a pyramid network, wherein the pyramid network includes: pyramid pooling layers, feature fusion layers, and a classifier; the pyramid network is used to fuse feature maps extracted by the backbone network at different depths and output a classification map through the classifier; and constructing a preset network structure based on the backbone network and the pyramid network.

[0012] Further, the step of iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the land cover recognition model includes: performing an initial iteration loop on the preset network structure using the initial label map and the target remote sensing image, obtaining the current label map output by the preset network structure after the last loop in the initial iteration loop and the iterated preset network structure; performing the next iteration loop on the iterated preset network structure using the current label map and the target remote sensing image, and updating the current label map during the iteration loop until the number of loops in the current iteration loop reaches the preset number of loops, obtaining the latest label map output by the preset network structure after the last loop in the current iteration loop and the current iterated preset network structure; and performing the next iteration loop on the current iterated preset network structure based on the latest label map and the target remote sensing image, until the preset number of iterations is reached to obtain the land cover recognition model.

[0013] Furthermore, the step of updating the current label map during the iterative loop includes: during the iterative loop, based on the current label map, calculating the confidence of the predicted label of each pixel output by the preset network structure after each loop; if the confidence is greater than the preset confidence threshold, replacing the label of the pixel in the current label map with the predicted label to generate the current label map for the next loop.

[0014] According to another aspect of the present invention, a land cover identification device for disaster response is also provided, comprising: a processing unit for acquiring remote sensing images of a disaster area and preprocessing the remote sensing images to obtain a target remote sensing image; a labeling unit for labeling the target remote sensing image to generate an initial label map, wherein the initial label map is used to characterize the initial label of each pixel on the target remote sensing image, and the initial label is used to characterize the land cover category of the pixel; an iteration unit for iterating a preset network structure based on the initial label map until the number of iterations reaches a preset number of iterations to obtain a land cover identification model, wherein the iteration of the preset network structure is achieved by continuously iterating the initial label map; and an identification unit for identifying the target remote sensing image using the land cover identification model to obtain a land cover classification map of the disaster area, wherein the land cover classification map is used to display the categories and location ranges of land cover in the disaster area.

[0015] Furthermore, the processing unit includes: a first processing module for performing radiometric calibration processing on the remote sensing image to obtain a first remote sensing image; a second processing module for performing atmospheric correction processing on the first remote sensing image to obtain a second remote sensing image; and a third processing module for performing orthorectification processing on the second remote sensing image to obtain a target remote sensing image.

[0016] Furthermore, the annotation unit includes: a first segmentation module, used to segment the target remote sensing image based on different scale parameters to obtain an image object set corresponding to each scale, wherein the image objects in the image object sets corresponding to different scales are of different sizes; a first selection module, used to select a preset number of image objects from the image object set for annotation for each image object set to obtain the annotation category of the image object; and a first classification module, used to classify the remaining image objects without annotation categories based on all image objects with annotation categories using a nearest neighbor classification strategy to generate an initial label map, wherein the initial label map is used to represent the initial category of each image object.

[0017] Further, the first classification module includes: a first extraction submodule, used to extract the feature vector of each image object, wherein the feature vector includes at least one of the following: spectral feature vector, texture feature vector, and shape feature vector; a first calculation submodule, used to calculate, for each remaining image object, a preset distance between the feature vector of the remaining image object and the feature vector of each preset image object, wherein the preset image object refers to an image object labeled with a labeled category; a first assignment submodule, used to assign the labeled category of the preset image object indicated by the minimum preset distance to the remaining image objects, thereby obtaining the initial category of the remaining image objects; and a first generation submodule, used to generate an initial label map based on the labeled categories of all preset image objects and the initial categories of all remaining image objects.

[0018] Furthermore, the ground feature recognition device also includes: a first construction module, used to construct a backbone network before iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the ground feature recognition model, wherein the backbone network includes: multiple convolutional structures, each convolutional structure corresponding to a deformable convolutional structure; the backbone network is used to extract feature maps from the input remote sensing image; a second construction module, used to construct a pyramid network, wherein the pyramid network includes: a pyramid pooling layer, a feature fusion layer, and a classifier; the pyramid network is used to fuse feature maps extracted by the backbone network at different depths and output a classification map through the classifier; a third construction module, used to construct a preset network structure based on the backbone network and the pyramid network.

[0019] Furthermore, the iterative unit includes: a first iterative module, used to perform an initial iterative loop on the preset network structure using the initial label map and the target remote sensing image, to obtain the current label map output by the preset network structure after the last loop in the initial iterative loop and the iterated preset network structure; a first update module, used to perform the next iterative loop on the iterated preset network structure using the current label map and the target remote sensing image, and update the current label map during the iterative loop until the number of loops in the current iterative loop reaches a preset number of loops, to obtain the latest label map output by the preset network structure after the last loop in the current iterative loop and the current iterated preset network structure; and a second iterative module, used to perform the next iterative loop on the current iterated preset network structure based on the latest label map and the target remote sensing image, until the number of iterations reaches a preset number of iterations, to obtain the land cover recognition model.

[0020] Furthermore, the first update module includes: a second calculation submodule, used to calculate the confidence of the predicted label of each pixel output by the preset network structure after each iteration based on the current label map during the iteration loop; and a second generation submodule, used to replace the label of the pixel in the current label map with the predicted label when the confidence is greater than the preset confidence threshold, and generate the current label map for use in the next loop.

[0021] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the ground feature identification method for disaster response described above.

[0022] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement any of the above-described methods for identifying ground features for disaster response.

[0023] In this invention, remote sensing images of disaster areas are acquired and preprocessed to obtain target remote sensing images. These target images are then labeled to generate an initial label map. Based on the initial label map, a preset network structure is iterated until the preset number of iterations is reached to obtain a ground feature recognition model. This model is then used to identify the target remote sensing images, resulting in a ground feature classification map of the disaster area. This solves the technical problem in related technologies where ground feature recognition cannot be performed quickly, leading to poor disaster response timeliness.

[0024] In this invention, multi-scale segmentation and nearest neighbor classification are used to generate an initial label map covering remote sensing images. Then, by iterating the initial label map, the preset network structure is continuously iterated. The resulting land cover recognition model can efficiently and accurately identify the category and location range of disaster-affected land cover on remote sensing images, improving its practicality and efficiency in disaster emergency response scenarios. It achieves the technical effect of quickly training the land cover recognition model with a small amount of manual annotation, thereby solving the technical problem that fully supervised methods rely heavily on large-scale, finely labeled datasets, resulting in slow response speed. Attached Figure Description

[0025] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0026] Figure 1 This is a flowchart of an optional feature identification method for disaster response according to an embodiment of the present invention;

[0027] Figure 2 This is a schematic diagram of an optional network structure according to an embodiment of the present invention;

[0028] Figure 3 This is a schematic diagram of an optional deep learning-based feature recognition process for rapid disaster response according to an embodiment of the present invention;

[0029] Figure 4 This is a schematic diagram of an optional feature identification device for disaster response according to an embodiment of the present invention;

[0030] Figure 5 This is a hardware structure block diagram of an electronic device (or mobile device) for a disaster response feature identification method according to an embodiment of the present invention. Detailed Implementation

[0031] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0032] It should be noted that the terms "first," "second," etc., used in this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0033] It should be noted that all related information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, stored data, and displayed data) collected and involved in this invention are information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, storage, use, processing, transmission, provision, disclosure, and application of this data comply with the relevant laws, regulations, and standards of the relevant regions, necessary confidentiality measures have been taken, and it does not violate public order and good morals. Corresponding operation entry points are provided for users to choose to authorize or refuse. For example, this system has an interface with relevant users or organizations. Before obtaining relevant information, a request to obtain the information needs to be sent to the aforementioned user or organization through the interface, and the relevant information is obtained only after receiving consent from the aforementioned user or organization.

[0034] This invention proposes a deep learning-based ground feature recognition method based on weakly supervised learning and iterative optimization. It is mainly applied after a geological disaster, using high-resolution aerial or satellite remote sensing images to quickly and automatically identify and segment disaster-affected ground features, providing decision support for emergency command, disaster assessment and rescue deployment.

[0035] To address the challenge of rapidly generating a high-precision land cover identification model with minimal manual annotations to meet the timeliness requirements of emergency response, this invention proposes a land cover identification method for rapid disaster response. This method combines image analysis methods (multi-scale segmentation and nearest neighbor classification) for generating coarse labels with a deep learning network for learning from and iteratively optimizing labels from the coarse labels.

[0036] To address the challenges of significantly reducing the model's reliance on manually labeled data and achieving rapid training and high-precision recognition with only a small amount of readily available prior knowledge, as well as the problem of generating a large number of effective training signals from a very small number of labels when directly using a small number of labeled samples for training, and how to generate a large number of imperfect but sufficient coarse labels to guide network learning, this invention proposes a method for generating full-image coarse labels from a very small number of labeled samples. This method involves first performing multi-scale segmentation of remote sensing images to form image objects, and then, based on a small number of manually labeled samples, automatically assigning categories to all unlabeled objects by calculating the feature distances between objects and using a nearest neighbor classifier.

[0037] Coarse labels generated from a small number of samples inevitably contain a large number of errors and blurred boundaries. Therefore, the problems of how to deal with noise and uncertainty in coarse labels, and how to design a deep learning network that can learn from noise and has self-correction and optimization capabilities, are the main issues. This invention proposes a label iterative optimization training mechanism that uses the high-confidence prediction results of the current model as "pseudo-labels" and continuously updates the training label set in subsequent training, thereby enabling the model to achieve self-improvement in performance in noisy label environments.

[0038] For disaster scenarios with complex conditions and significant differences in the scale of affected landforms, this invention proposes a specific network architecture to address the challenges of improving the accuracy of multi-scale disaster target identification and ensuring robust identification capabilities for landforms at different scales. It employs InternImage (a computer vision model) as the backbone network to enhance the ability to model complex contexts, and combines it with a UPerNet (Unified Perceptual Parsing Network, a deep learning network architecture for image segmentation) segmentation head to achieve accurate segmentation of multi-scale targets. This combination is particularly suitable for remote sensing disaster scenarios with complex and variable landform morphology.

[0039] Compared to fully supervised methods, the method proposed in this invention effectively reduces the workload of manual annotation, shortening model readiness time from several days to several hours. Furthermore, using only a small number of labeled samples (less than 15% of the total area), the final accuracy of ground feature identification can reach or approach the level of fully supervised training methods. In addition, the proposed method framework is universal and can be easily transferred to image analysis tasks involving different types of disasters (such as floods and fires) and different sensors (such as SAR (Synthetic Aperture Radar)).

[0040] The present invention will now be described in detail with reference to various embodiments.

[0041] Example 1

[0042] According to an embodiment of the present invention, an embodiment of a ground feature identification method for disaster response is provided. It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions. Furthermore, although a logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order than that shown here.

[0043] Figure 1 This is a flowchart of an optional feature identification method for disaster response according to an embodiment of the present invention, such as... Figure 1 As shown, the method includes the following steps:

[0044] Step S101: Acquire remote sensing images of the disaster area and preprocess the remote sensing images to obtain target remote sensing images.

[0045] In this embodiment of the invention, after a disaster occurs in a certain area, remote sensing image data of the disaster area can be collected in a timely manner using aerial drones or high-resolution satellites. Then, these images undergo a series of standard preprocessing measures, including radiometric calibration, atmospheric correction, and orthorectification, to eliminate errors caused by sensor and environmental factors, ensuring that the true surface reflectance information reflected by the images is not affected by sensor characteristics and weather conditions, thereby obtaining the target remote sensing image.

[0046] Step S102: Label the target remote sensing image to generate an initial label map, wherein the initial label map is used to characterize the initial label of each pixel on the target remote sensing image, and the initial label is used to characterize the land cover category of the pixel.

[0047] In this embodiment of the invention, initial training labels (i.e., initial label maps) covering the entire image can be quickly generated using a minimal amount of manual annotation. On the preprocessed image, a multi-scale segmentation algorithm can be applied to divide the image into several homogeneous, non-overlapping image objects, or superpixels, transforming the basic unit of analysis from "pixels" into "objects" with certain semantic information. Then, representative image objects can be manually labeled, such as "collapsed buildings" and "uncollapsed buildings." Based on these few annotations, preliminary category labels are assigned to unlabeled image objects, generating an initial label map (i.e., coarse labels) covering the entire image. Although these labels are relatively coarse, they provide preliminary supervisory information, giving the deep learning network a basic training starting point. This significantly reduces the workload of manual annotation and improves the efficiency of the data preparation stage.

[0048] Step S103: Iterate the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the land feature recognition model. The iteration of the preset network structure is achieved by continuously iterating the initial label map.

[0049] In this embodiment of the invention, the preset network structure refers to a deep learning model framework composed of an InternImage backbone network and a UPerNet segmentation head. Based on the initial label map, the network structure is trained for the first time (called initialization training), followed by an iterative optimization phase. The network structure uses its own prediction results (when the confidence of the result exceeds a preset threshold) as new training signals (i.e., "pseudo-labels"), and merges or replaces them with the previous label map to form a higher-quality label set. This process is repeated, and the number of iterations can be preset or dynamically terminated based on the model performance convergence criterion. Through continuous iteration of the label set and retraining, the network structure gradually corrects errors in the initial labels and optimizes network weights until satisfactory recognition performance is achieved, ultimately obtaining the land cover recognition model.

[0050] Step S104: Use a land cover recognition model to identify the target remote sensing image and obtain a land cover classification map of the disaster area. The land cover classification map is used to show the category and location range of land cover in the disaster area.

[0051] In this embodiment of the invention, after the model is trained, it is deployed onto the remote sensing image of the disaster area to be analyzed for forward inference. The model predicts the land cover category of each pixel, forming a detailed "land cover classification map." This classification map not only accurately marks the location and extent of disaster-affected land cover, but also distinguishes different types of land cover, such as buildings, roads, vegetation, and water bodies, providing intuitive visual evidence and quantitative data support for post-disaster assessment and rescue operations. In this way, it can quickly provide scientific and accurate decision support for disaster response work.

[0052] In some alternative embodiments, the precise location and extent of features in the disaster area shown on the feature classification map can be further quantified and statistically analyzed to output a disaster assessment report.

[0053] In summary, multi-scale segmentation and nearest neighbor classification can be used to generate initial label maps covering remote sensing images. Then, by iterating the initial label maps, the preset network structure can be continuously iterated. The resulting land cover recognition model can efficiently and accurately identify the category and location range of disaster-affected land cover on remote sensing images, improving its practicality and efficiency in disaster emergency response scenarios. It achieves the technical effect of rapidly training the land cover recognition model with a small amount of manual annotation, thus solving the technical problem of slow response speed caused by the heavy reliance of fully supervised methods on large-scale, finely labeled datasets.

[0054] To improve the accuracy of preprocessing remote sensing images, in the land cover identification method for disaster response provided in Embodiment 1 of this application, the remote sensing images are radiometrically calibrated to obtain a first remote sensing image; the first remote sensing image is atmospherically corrected to obtain a second remote sensing image; and the second remote sensing image is orthorectified to obtain a target remote sensing image.

[0055] In this embodiment of the invention, to ensure the recognition accuracy of the deep learning model in disaster response, the acquired raw remote sensing images can be radiometrically calibrated. Radiometric calibration refers to the process of converting digital values ​​on an image into physical radiance or reflectivity, which can compensate for the difference between the energy received by the sensor and the actual energy emitted or reflected by the ground surface, and eliminate image distortion caused by factors such as fluctuations in sensor sensitivity and changes in lighting conditions.

[0056] Then, atmospheric correction is performed on the first remote sensing image obtained after radiometric calibration. The purpose of atmospheric correction is to remove the influence of atmospheric scattering and absorption effects on the measured surface reflectance. Due to differences in atmospheric composition (such as water vapor and aerosols) and state, the same ground feature may exhibit different brightness or color in remote sensing images at different times, which can interfere with the learning process of deep learning networks. Through atmospheric correction, these biases can be calibrated to ensure the consistency and comparability of ground feature attributes, thereby improving the model's generalization ability and recognition accuracy.

[0057] Next, the atmospherically corrected second remote sensing image undergoes orthorectification. Orthorectification is the process of converting oblique aerial or satellite imagery into an orthophoto. Orthorectification can eliminate geometric distortions caused by changes in flight altitude, angle, and terrain undulations, ensuring that each pixel in the image corresponds to a specific location on the ground.

[0058] In this embodiment, preprocessing steps such as radiometric calibration, atmospheric correction, and orthorectification transform the original remote sensing image into a target remote sensing image. This series of preprocessing operations not only improves the image quality but also standardizes the image data, creating favorable conditions for the training and application of deep learning-based ground feature recognition models.

[0059] To improve the accuracy of generating initial label maps, in the land cover identification method for disaster response provided in Embodiment 1 of this application, the target remote sensing image is segmented based on different scale parameters to obtain an image object set corresponding to each scale. The image objects in the image object sets corresponding to different scales are of different sizes. For each image object set, a preset number of image objects are selected from the image object set for labeling to obtain the label category of the image object. Based on all image objects labeled with label categories, the remaining image objects without label categories are classified using a nearest neighbor classification strategy to generate an initial label map. The initial label map is used to represent the initial category of each image object.

[0060] In this embodiment of the invention, a multi-scale segmentation algorithm can be used to segment the image into several homogeneous, non-overlapping image objects, or superpixels. By setting different scale parameters, objects of different sizes can be obtained to accommodate ground features of different sizes. In this way, the basic unit of analysis can be transformed from a "pixel" to an "object" with certain semantic information, avoiding the information redundancy and computational complexity that pixel-level processing may bring. Specifically, a series of scale parameters can be set, which control the size of the image objects. Generally, small-scale parameters will generate more and smaller image objects, while large-scale parameters will generate fewer and larger objects. Choosing an appropriate set of scale parameters is to ensure that the characteristics of ground features at different scales are captured, including small buildings, roads, vegetation, and large areas such as cities and lakes. Then, according to these scale parameters, the target remote sensing image is segmented to obtain the image object set corresponding to each scale.

[0061] In this embodiment of the invention, operators can select and label image objects representing different land cover types based on disaster response needs and the characteristics of remote sensing imagery. This labeling can be very intuitive; for example, a land cover category can be assigned to an image object simply by clicking or outlining it. For instance, several typical "collapsed building" areas and "non-collapsed building" areas can be selected. Here, "preset quantity" refers to the number of image objects that the operator needs to manually label. The labeling category refers to the land cover type represented by the image object, such as "building," "road," or "water body."

[0062] In this embodiment of the invention, a nearest neighbor classification strategy can be used to classify the remaining image objects that are not labeled with a category, thereby obtaining an initial category for each image object and thus an initial label map. For example, for each unlabeled image object, its feature vector is extracted, which may include spectral features, texture features, shape features, etc. Then, the distance between this feature vector and the feature vectors of all labeled samples is calculated, and the sample with the smallest distance is selected as the nearest neighbor, thereby obtaining the preliminary category of the image object. This process is applied to the entire image to generate an initial label map covering the entire image. Although this classification may not be accurate enough due to the limitations of the feature space, it provides the necessary training data for subsequent deep learning model iterations. Furthermore, the nearest neighbor classification strategy can quickly generate training labels covering the entire image, reducing the burden of manual labeling, while allowing the model to learn from these preliminary labels and gradually correct and optimize its recognition ability.

[0063] In this embodiment, rapid generation of initial label maps from raw remote sensing imagery is achieved. Through multi-scale segmentation and nearest-neighbor classification strategies, not only is the workload of manual annotation reduced, but the model is also ensured to learn rich feature information from representative ground cover samples, laying the foundation for subsequent iterative optimization and high-precision identification.

[0064] To further improve the accuracy of generating the initial label map, in the land cover identification method for disaster response provided in Embodiment 1 of this application, the feature vector of each image object is extracted, wherein the feature vector includes at least one of the following: spectral feature vector, texture feature vector, and shape feature vector; for each remaining image object, a preset distance is calculated between the feature vector of the remaining image object and the feature vector of each preset image object, wherein the preset image object refers to an image object labeled with a label category; the label category of the preset image object indicated by the minimum preset distance is assigned to the remaining image objects to obtain the initial category of the remaining image objects; an initial label map is generated based on the label categories of all preset image objects and the initial categories of all remaining image objects.

[0065] In this embodiment of the invention, for each segmented image object, predefined methods (such as statistical analysis, texture analysis algorithms, shape analysis algorithms, etc.) can be used to calculate and store its feature vector. For example, the spectral feature vector can be obtained by calculating the average gray value of the image object in each band; the texture feature vector can be extracted using texture analysis tools such as the gray-level co-occurrence matrix; and the geometric attributes of the object can be calculated using its boundary pixels to obtain the shape feature vector. Here, the feature vector of the image object is a mathematical representation of its attributes, which may include spectral feature vectors, texture feature vectors, shape feature vectors, etc. Spectral feature vectors typically reflect the average reflectance of the image object in different bands (such as red, green, blue, near-infrared, etc.); texture feature vectors describe the internal texture patterns of the image object, such as uniformity, complexity, etc.; and shape feature vectors encode the geometric attributes of the object, such as area, perimeter, aspect ratio, etc.

[0066] In this embodiment of the invention, for each unlabeled image object, its feature vector is compared with the feature vectors of all preset image objects, and a preset distance (such as Euclidean distance) is calculated between them. For example, this is achieved through vector subtraction and square root operations, ultimately obtaining a distance matrix to record the distance between the unlabeled image object and each preset image object. Then, according to the nearest neighbor principle, the preset image object corresponding to the minimum distance is found, and its labeling category is assigned to the unlabeled image object (i.e., the remaining image objects). In this way, each unlabeled image object is assigned an initial category that is closest to the preset image object based on its similarity. Afterwards, the categories of all image objects, whether directly labeled or inferred through the nearest neighbor strategy, are integrated into a label map that covers the entire image, thereby forming a coarse label map (i.e., the initial label map) covering the entire training area. Here, the coarse label map is a label map with good internal consistency, but the object boundaries may be inaccurate and there may be some misclassification.

[0067] In this embodiment, the manual annotation process is simplified while fully utilizing the comprehensive feature information of image objects, providing a fast startup path for deep learning models. The generation of the initial label map not only accelerates the training preparation of the model but also makes it possible to improve recognition accuracy through model self-optimization under limited labeled data conditions, providing strong technical support for rapid disaster response.

[0068] To accurately construct the preset network structure, in the land cover recognition method for disaster response provided in Embodiment 1 of this application, before iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the land cover recognition model, a backbone network is constructed. The backbone network includes multiple convolutional structures, each corresponding to a deformable convolutional structure. The backbone network is used to extract feature maps from the input remote sensing image. A pyramid network is constructed, including a pyramid pooling layer, a feature fusion layer, and a classifier. The pyramid network is used to fuse feature maps extracted from the backbone network at different depths and output a classification map through the classifier. Based on the backbone network and the pyramid network, the preset network structure is constructed.

[0069] In this embodiment of the invention, the model structure (i.e., the preset network structure) for ground feature recognition may include: a backbone network and a segmentation head (such as a pyramid network). The backbone network may use InternImage. InternImage is a large-kernel convolutional network that dynamically adjusts its receptive field through deformable convolution, enabling it to better capture long-range dependencies and complex contextual information, making it very suitable for handling disaster scenarios with mixed ground features and diverse shapes. The segmentation head may use UPerNet. The pyramid network (Feature Pyramid Network, FPN) structure in UPerNet can effectively fuse feature maps (high-level semantic information and low-level detail information) extracted from the backbone network at different depths, thereby achieving accurate segmentation of multi-scale targets.

[0070] In this embodiment of the invention, InternImage is used as the backbone network architecture, which is a network design based on large kernel convolution and deformable convolution. The backbone network includes multiple convolutional structures, each corresponding to a deformable convolutional structure. The backbone network is used to extract feature maps from the input remote sensing image. In this way, it can adaptively determine which locations to focus on and the degree of focus based on the content of the input image. It is particularly suitable for handling problems with complex distribution and varying scales of ground features in remote sensing images, and can more effectively capture the details and contextual information of ground features.

[0071] In this embodiment of the invention, a UPerNet pyramid network can be connected after the InternImage backbone network. The pyramid network is used to fuse feature maps extracted from the backbone network at different depths and output a classification map through a classifier. The pyramid network includes: pyramid pooling layers, feature fusion layers, and a classifier. The pyramid pooling layers can perform pooling operations on the feature maps at different scales, thereby obtaining a series of feature representations with different resolutions, covering land cover information from coarse to detailed. The feature fusion layer fuses the multi-scale features generated by the pyramid pooling layers, constructing a feature pyramid through cross-layer connections and upsampling operations, which contains rich features from low to high levels. The classifier is located at the top of the pyramid network and is responsible for transforming the fused feature map into the final classification map. By classifying pixels one by one, each pixel is assigned a probability of belonging to a certain land cover category. Then, the constructed InternImage backbone network is combined with the UPerNet pyramid network to form a complete deep learning model (i.e., the preset network structure). This model first extracts the low-level and high-level features of the input remote sensing image through a backbone network, then performs feature fusion through a pyramid network, and finally outputs the final classification map by the classifier at the top of the pyramid network.

[0072] In this embodiment, a powerful and efficient deep learning model was obtained, capable of automatically identifying and classifying various ground features from remote sensing imagery, maintaining high accuracy even in drastically changing post-disaster environments. The combination of the backbone network and the pyramid network not only enhances the model's understanding of ground feature details and contextual relationships but also strengthens its ability to handle multi-scale targets. This model structure design opens up new avenues for rapid analysis and emergency applications of remote sensing imagery, particularly in disaster response scenarios with limited resources and tight deadlines, significantly improving the response speed and reliability of decision support systems.

[0073] Figure 2 This is a schematic diagram of an optional network structure according to an embodiment of the present invention, such as... Figure 2 As shown, the network structure includes InternImage and UPerNet. InternImage comprises multiple 3×3 convolutions, each followed by a different deformable convolutional stage (e.g., the first 3×3 convolution is followed by stage 1×5, the second by stage 2×5, the third by stage 3×22, and the fourth by stage 4×5). Each stage i (stages 1 to 4) contains a structure composed of multiple deformable convolutions and layer normalization. UPerNet includes pyramid pooling, a feature pyramid network, feature fusion, 3×3 convolutions, and a classifier.

[0074] exist Figure 2 In the network structure shown, remote sensing images can first be processed by InternImage and then input into UPerNet. After processing through pyramid pooling, feature pyramid network, feature fusion, etc., the fused feature map is obtained. Then, the classification result is output through 3×3 convolution and classifier.

[0075] To accurately iterate the preset network structure, in the land cover identification method for disaster response provided in Embodiment 1 of this application, an initial label map and target remote sensing image are used to perform an initial iteration loop on the preset network structure, resulting in the current label map output by the preset network structure after the last loop in the initial iteration loop and the iterated preset network structure. The current label map and target remote sensing image are then used to perform the next iteration loop on the iterated preset network structure, updating the current label map during the iteration loop until the number of iterations in the current iteration loop reaches a preset number of iterations, resulting in the latest label map output by the preset network structure after the last loop in the current iteration loop and the currently iterated preset network structure. Based on the latest label map and target remote sensing image, the currently iterated preset network structure is then used to perform the next iteration loop until the number of iterations reaches a preset number of iterations, thus obtaining the land cover identification model.

[0076] In this embodiment of the invention, a coarse label map (denoted as L0) can be used as supervision information for initial training of the network. Since the labels are coarse, the resulting model (denoted as M0) has limited performance. Then, an iterative loop is performed: the current model M... i (i is the current iteration number) is applied to the original image to obtain a predicted probability map. For each pixel, if its prediction confidence is higher than a preset threshold θ, the prediction is considered reliable and used as a new training label. The generated pseudo-labels are then fused with or replaced with the labels from the previous round to form a new, higher-quality label set L. i+1 Using the new tag set L i+1 The network is retrained to obtain a more powerful model M. i+1 Repeat this loop until the model performance converges or the preset number of iterations is reached. Through this process, the network uses its high-confidence output to continuously refine the initial coarse labels, thereby gradually improving its recognition accuracy.

[0077] Specifically, at the beginning of the first iteration, the initial label map and the target remote sensing image are input into the constructed deep learning network. The network extracts image features through the InternImage backbone network, then performs feature fusion and classification through the pyramid network, outputting a predicted current label map. Subsequently, based on the difference between the current label map and the initial label map, the network's weight parameters are updated, completing one iteration. This process is repeated until the last iteration, at which point the resulting current label map becomes the input for the next iteration.

[0078] In each iteration, the current labeled image and the target remote sensing image are used as input to run a pre-defined network structure, outputting a new predicted labeled image. Then, based on a confidence comparison between the new predicted labeled image and the current labeled image, higher-precision pseudo-labels are generated. These pseudo-labels are used to update the model's training, further improving its recognition capabilities. This process continues until the preset number of iterations is reached, at which point the latest labeled image is used to begin the next iteration. After reaching the preset number of iterations, the iteration cycle is executed again based on the latest predicted labeled image and the target remote sensing image. This process may involve more optimization strategies, such as learning rate decay and regularization, to prevent overfitting. The iteration cycle continues until the preset number of iterations is reached. After the final iteration, the resulting model is the final land cover recognition model, capable of automatically identifying and classifying land covers in remote sensing images with high accuracy.

[0079] In this embodiment, through continuous feedback and optimization, the deep learning model's dependence on finely labeled data is significantly reduced. This allows the model to rapidly improve recognition accuracy even with relatively coarse initial labels, achieving results close to those of fully supervised training methods. It is particularly suitable for rapid disaster response scenarios, generating high-quality land cover recognition and classification results within a limited time, thus providing timely and accurate support for emergency decision-making. Through iterative iteration, the model demonstrates excellent adaptability and generalization ability in practical applications, effectively addressing the recognition needs of different types of disaster scenarios.

[0080] To improve the accuracy of updating the current label map, in the land cover identification method for disaster response provided in Embodiment 1 of this application, during the iterative loop, based on the current label map, the confidence level of the predicted label of each pixel output by the preset network structure after each loop is calculated; if the confidence level is greater than the preset confidence threshold, the predicted label is used to replace the label of the pixel in the current label map, and the current label map used in the next loop is generated.

[0081] In this embodiment of the invention, at the end of each iteration, the deep learning network outputs a predicted current label map. Based on this label map, the confidence level of each pixel being classified into a certain land cover category can be calculated. Specifically, the network outputs a classification probability map, reads the probability of each pixel being classified into each land cover category, and then determines the category with the highest probability (highest confidence) as the predicted label for that pixel. The confidence level is calculated by taking this highest probability value. After obtaining the predicted confidence level for each pixel, it is checked whether the confidence level of each pixel exceeds a preset confidence threshold θ. If the confidence level of a pixel does exceed the threshold, the old label for that pixel in the current label map is replaced with the predicted label, thus generating an updated current label map. This new label map will be used for training in the next iteration.

[0082] For example, the confidence level is calculated using L0 (rough label map) and L... n IoU (Intersection over Union) of the labels in the nth iteration:

[0083]

[0084] Where θ is the preset confidence threshold, and L i L represents the label for the i-th iteration. i+1 This represents the label for the (i+1)th iteration.

[0085] In this embodiment, through confidence calculation and label replacement mechanisms, the model can self-feedback and optimize during the learning process, gradually improving its accuracy in recognizing ground features in remote sensing imagery. The pre-set confidence threshold ensures both the quality of the model's output and the efficiency of model training. As the number of iterations increases, the model's predictions become increasingly closer to reality, and the quality of the final label map continuously improves until the model converges, achieving satisfactory recognition performance.

[0086] Figure 3 This is a schematic diagram of an optional deep learning-based feature recognition process for rapid disaster response according to an embodiment of the present invention, as shown below. Figure 3 As shown, a rapid ground feature identification is achieved through a coarse label generation + label iterative optimization network. The specific process is as follows:

[0087] (1) Perform multi-scale segmentation on the high-resolution remote sensing images of the disaster area that have been acquired;

[0088] (2) Manually label the land cover categories of the segmented objects in small quantities;

[0089] (3) For unclassified objects, the nearest neighbor classification is used to determine the category in order to obtain the initial label;

[0090] (4) Train the deep learning network based on the initial dataset (including initial labels);

[0091] (5) Calculate the IoU of the prediction results based on the prediction results and the initial labels;

[0092] (6) If IoU>θ, replace the initial label with the prediction result; otherwise, use the initial label to obtain the iterative dataset.

[0093] (7) Train the deep learning network continuously based on the iterative dataset until the model converges / reaches the required number of iterations.

[0094] In this embodiment of the invention, multi-scale segmentation and automated coarse label generation technology based on nearest neighbors significantly reduce data preparation time, making it possible to rapidly deploy deep learning models within the precious post-disaster timeframe and solving the fundamental timeliness bottleneck in current emergency response. Furthermore, the iterative label optimization mechanism cleverly utilizes the network's own learning capabilities to overcome noise issues in the initial labels. In addition, the InternImage+UPerNet architecture is better suited for rapid disaster response scenarios in terms of feature extraction and multi-scale processing capabilities. InternImage's large receptive field allows it to better understand the overall layout of the disaster scene, while UpperNet's multi-level feature fusion ensures accurate identification of disaster-affected features of different sizes. Therefore, the model in this embodiment exhibits higher recognition accuracy and robustness when facing complex and diverse disaster scenarios.

[0095] The following is a detailed description with reference to another embodiment.

[0096] Example 2

[0097] The ground feature identification device for disaster response provided in this embodiment includes multiple implementation units, each of which corresponds to a specific implementation step in Embodiment 1 above.

[0098] Figure 4 This is a schematic diagram of an optional feature identification device for disaster response according to an embodiment of the present invention, such as... Figure 4 As shown, the feature identification device may include: a processing unit 40, a labeling unit 41, an iteration unit 42, and an identification unit 43.

[0099] The processing unit 40 is used to acquire remote sensing images of the disaster area and preprocess the remote sensing images to obtain target remote sensing images.

[0100] Labeling unit 41 is used to label the target remote sensing image and generate an initial label map, wherein the initial label map is used to represent the initial label of each pixel on the target remote sensing image, and the initial label is used to represent the land cover category of the pixel;

[0101] Iteration unit 42 is used to iterate the preset network structure based on the initial label map until the number of iterations reaches the preset number of iterations to obtain the land feature recognition model. The iteration of the preset network structure is achieved by continuously iterating the initial label map.

[0102] The identification unit 43 is used to identify the target remote sensing image using a ground feature identification model to obtain a ground feature classification map of the disaster area. The ground feature classification map is used to show the category and location range of ground features in the disaster area.

[0103] The aforementioned land cover recognition device can generate an initial label map covering remote sensing images by using multi-scale segmentation and nearest neighbor classification. Then, by iterating the initial label map, the device continuously iterates the preset network structure. The resulting land cover recognition model can efficiently and accurately identify the category and location range of disaster-affected land cover on remote sensing images, improving its practicality and efficiency in disaster emergency response scenarios. It achieves the technical effect of rapidly training the land cover recognition model with a small amount of manual annotation, thereby solving the technical problem that fully supervised methods rely heavily on large-scale, finely labeled datasets, resulting in slow response speeds.

[0104] Optionally, the processing unit 40 includes: a first processing module for performing radiometric calibration processing on the remote sensing image to obtain a first remote sensing image; a second processing module for performing atmospheric correction processing on the first remote sensing image to obtain a second remote sensing image; and a third processing module for performing orthorectification processing on the second remote sensing image to obtain a target remote sensing image.

[0105] Optionally, the annotation unit 41 includes: a first segmentation module, used to segment the target remote sensing image based on different scale parameters to obtain an image object set corresponding to each scale, wherein the image objects in the image object sets corresponding to different scales are of different sizes; a first selection module, used to select a preset number of image objects from the image object set for annotation for each image object set to obtain the annotation category of the image object; and a first classification module, used to classify the remaining image objects without annotation categories based on all image objects with annotation categories using a nearest neighbor classification strategy to generate an initial label map, wherein the initial label map is used to represent the initial category of each image object.

[0106] Optionally, the first classification module includes: a first extraction submodule, used to extract the feature vector of each image object, wherein the feature vector includes at least one of the following: spectral feature vector, texture feature vector, and shape feature vector; a first calculation submodule, used to calculate, for each remaining image object, a preset distance between the feature vector of the remaining image object and the feature vector of each preset image object, wherein the preset image object refers to an image object labeled with a labeled category; a first assignment submodule, used to assign the labeled category of the preset image object indicated by the minimum preset distance to the remaining image objects, thereby obtaining the initial category of the remaining image objects; and a first generation submodule, used to generate an initial label map based on the labeled categories of all preset image objects and the initial categories of all remaining image objects.

[0107] Optionally, the ground feature recognition device further includes: a first construction module, used to construct a backbone network before iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the ground feature recognition model, wherein the backbone network includes: multiple convolutional structures, each convolutional structure corresponding to a deformable convolutional structure; the backbone network is used to extract feature maps from the input remote sensing image; a second construction module, used to construct a pyramid network, wherein the pyramid network includes: a pyramid pooling layer, a feature fusion layer, and a classifier; the pyramid network is used to fuse feature maps extracted by the backbone network at different depths and output a classification map through the classifier; a third construction module, used to construct a preset network structure based on the backbone network and the pyramid network.

[0108] Optionally, the iteration unit 42 includes: a first iteration module, used to perform an initial iteration loop on the preset network structure using the initial label map and the target remote sensing image, to obtain the current label map output by the preset network structure after the last loop in the initial iteration loop and the iterated preset network structure; a first update module, used to perform the next iteration loop on the iterated preset network structure using the current label map and the target remote sensing image, and update the current label map during the iteration loop until the number of loops in the current iteration loop reaches a preset number of loops, to obtain the latest label map output by the preset network structure after the last loop in the current iteration loop and the current iterated preset network structure; and a second iteration module, used to perform the next iteration loop on the current iterated preset network structure based on the latest label map and the target remote sensing image, until the number of iterations reaches a preset number of iterations, to obtain the land cover recognition model.

[0109] Optionally, the first update module includes: a second calculation submodule, used to calculate the confidence of the predicted label of each pixel output by the preset network structure after each iteration based on the current label map during the iteration loop; and a second generation submodule, used to replace the label of the pixel in the current label map with the predicted label when the confidence is greater than a preset confidence threshold, and generate the current label map for use in the next loop.

[0110] The aforementioned feature identification device may also include a processor and a memory. The aforementioned processing unit 40, labeling unit 41, iteration unit 42, identification unit 43, etc., are all stored in the memory as program units, and the processor executes the aforementioned program units stored in the memory to realize the corresponding functions.

[0111] The aforementioned processor contains a kernel, which retrieves the corresponding program units from memory. One or more kernels can be configured, and by adjusting kernel parameters, a land cover recognition model can be used to identify targets in remote sensing images, resulting in a land cover classification map of the disaster area.

[0112] The aforementioned memory may include non-permanent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM, and the memory includes at least one memory chip.

[0113] The present invention also provides a computer program product, which, when executed on a data processing device, is suitable for executing an initialization program having the following method steps: acquiring remote sensing images of a disaster area, preprocessing the remote sensing images to obtain target remote sensing images, labeling the target remote sensing images to generate an initial label map, iterating a preset network structure based on the initial label map until the number of iterations reaches a preset number of iterations to obtain a ground feature recognition model, and using the ground feature recognition model to identify the target remote sensing images to obtain a ground feature classification map of the disaster area.

[0114] According to another aspect of the present invention, a computer program product is also provided, including a non-volatile computer-readable storage medium storing a computer program, which, when executed by a processor, implements the ground feature identification method for disaster response described above.

[0115] According to another aspect of the present invention, an electronic device is also provided, including one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the above-described method for identifying ground features for disaster response.

[0116] Figure 5This is a hardware structure block diagram of an electronic device (or mobile device) for a disaster response feature identification method according to an embodiment of the present invention. Figure 5 As shown, an electronic device may include one or more processors (e.g., Figure 5 The processors 502a, 502b, ..., 502n, etc., may include, but are not limited to, processing devices such as microprocessors (MCUs) or programmable logic devices (FPGAs), and a memory 504 for storing data. In addition, it may include: a display, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports in the I / O interface), a network interface, a keyboard, a power supply, and / or a camera. Those skilled in the art will understand that... Figure 5 The structure shown is for illustrative purposes only and does not limit the structure of the electronic device described above. For example, the electronic device may also include components that are more... Figure 5 The more or fewer components shown, or having the same Figure 5 The different configurations shown.

[0117] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0118] The embodiments or examples disclosed herein are not exhaustive, but merely illustrative of some embodiments or examples, and are not intended to limit the scope of protection of this disclosure. Unless otherwise specified, each step in a particular embodiment or example can be implemented as an independent embodiment, and the steps can be arbitrarily combined. For example, a solution after removing some steps in a particular embodiment or example can also be implemented as an independent embodiment, and the order of the steps in a particular embodiment or example can be arbitrarily interchanged. Furthermore, optional methods or examples in a particular embodiment or example can be arbitrarily combined; moreover, embodiments or examples can be arbitrarily combined. For example, some or all steps of different embodiments or examples can be arbitrarily combined, and a particular embodiment or example can be arbitrarily combined with optional methods or examples of other embodiments or examples.

[0119] In the above embodiments of the present invention, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0120] In the several embodiments provided by this invention, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling, direct coupling, or communication connection can be through some interfaces; the indirect coupling or communication connection of units or modules can be electrical or other forms.

[0121] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0122] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0123] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0124] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for identifying ground features in disaster response, characterized in that, include: Acquire remote sensing images of the disaster area and preprocess the remote sensing images to obtain target remote sensing images; The target remote sensing image is labeled to generate an initial label map, wherein the initial label map is used to characterize the initial label of each pixel on the target remote sensing image, and the initial label is used to characterize the land cover category of the pixel; The preset network structure is iterated based on the initial label map until the preset number of iterations is reached, thereby obtaining the land feature recognition model. The iteration of the preset network structure is achieved by continuously iterating the initial label map. The land cover identification model is used to identify the target remote sensing image to obtain a land cover classification map of the disaster area, wherein the land cover classification map is used to show the category and location range of land cover in the disaster area.

2. The feature identification method according to claim 1, characterized in that, The steps of preprocessing the remote sensing image to obtain the target remote sensing image include: The remote sensing image is radiometrically calibrated to obtain the first remote sensing image; The first remote sensing image is subjected to atmospheric correction to obtain the second remote sensing image; The second remote sensing image is subjected to orthorectification to obtain the target remote sensing image.

3. The feature identification method according to claim 1, characterized in that, The steps of annotating the target remote sensing image and generating an initial label map include: Based on different scale parameters, the target remote sensing image is segmented to obtain an image object set corresponding to each scale, wherein the size of the image objects in the image object set corresponding to different scales is different; For each set of image objects, a preset number of image objects are selected from the set of image objects for annotation to obtain the annotation category of the image objects; Based on all the image objects labeled with the labeled categories, the remaining image objects that are not labeled with the labeled categories are classified using a nearest neighbor classification strategy to generate the initial label map, wherein the initial label map is used to characterize the initial category of each image object.

4. The feature identification method according to claim 3, characterized in that, The step of generating the initial label map by classifying the remaining image objects that are not labeled with the specified label categories based on all image objects labeled with the specified label categories using a nearest neighbor classification strategy includes: Extract the feature vector for each of the image objects, wherein the feature vector includes at least one of the following: spectral feature vector, texture feature vector, and shape feature vector; For each remaining image object, a preset distance is calculated between the feature vector of the remaining image object and the feature vector of each preset image object, wherein the preset image object refers to an image object labeled with the label category; The label category of the preset image object indicated by the minimum preset distance is assigned to the remaining image object to obtain the initial category of the remaining image object; The initial label map is generated based on the label categories of all the preset image objects and the initial categories of all the remaining image objects.

5. The feature identification method according to claim 1, characterized in that, Before iterating the preset network structure based on the initial label map until the preset number of iterations is reached and the land cover recognition model is obtained, the process further includes: A backbone network is constructed, comprising: multiple convolutional structures, each of which corresponds to a deformable convolutional structure; the backbone network is used to extract feature maps from the input remote sensing image. A pyramid network is constructed, comprising: a pyramid pooling layer, a feature fusion layer, and a classifier; the pyramid network is used to fuse the feature maps extracted from the backbone network at different depths, and outputs a classification map through the classifier; Based on the backbone network and the pyramid network, the preset network structure is constructed.

6. The feature identification method according to claim 1, characterized in that, The steps of iterating the preset network structure based on the initial label map until the preset number of iterations is reached to obtain the land cover recognition model include: The initial label map and the target remote sensing image are used to perform an initial iteration loop on the preset network structure to obtain the current label map output by the preset network structure after the last loop in the initial iteration loop and the iterated preset network structure. The current label map and the target remote sensing image are used to perform the next iteration loop on the preset network structure after iteration. During the iteration loop, the current label map is updated until the number of iterations in the current iteration loop reaches the preset number of iterations. The latest label map output by the preset network structure after the last iteration in the current iteration loop and the preset network structure after the current iteration are obtained. Based on the latest label map and the target remote sensing image, the preset network structure after the current iteration is iterated again until the preset number of iterations is reached, and the land feature recognition model is obtained.

7. The feature identification method according to claim 6, characterized in that, The step of updating the current label graph during the iterative loop includes: During the iterative loop, based on the current label map, the confidence level of the predicted label for each pixel output by the preset network structure after each loop is calculated; If the confidence level is greater than a preset confidence threshold, the predicted label is used to replace the label of the pixel in the current label map, and a current label map for the next cycle is generated.

8. A feature identification device for disaster response, characterized in that, include: The processing unit is used to acquire remote sensing images of the disaster area and preprocess the remote sensing images to obtain target remote sensing images; A labeling unit is used to label the target remote sensing image and generate an initial label map, wherein the initial label map is used to characterize the initial label of each pixel on the target remote sensing image, and the initial label is used to characterize the land cover category of the pixel; An iterative unit is used to iterate the preset network structure based on the initial label map until the number of iterations reaches the preset number of iterations to obtain the land feature recognition model. The iteration of the preset network structure is achieved by continuously iterating the initial label map. The identification unit is used to identify the target remote sensing image using the land feature identification model to obtain a land feature classification map of the disaster area, wherein the land feature classification map is used to display the category and location range of land features in the disaster area.

9. A computer program product, characterized in that, The method includes a non-volatile computer-readable storage medium storing a computer program that, when executed by a processor, implements the feature identification method for disaster response as described in any one of claims 1 to 7.

10. An electronic device, characterized in that, It includes one or more processors and a memory, the memory being used to store one or more programs, wherein when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the feature identification method for disaster response as described in any one of claims 1 to 7.

Images